def subsample(inputs, factor, scope=None):
"""Subsamples the input along the spatial dimensions.
Args:
inputs: A `Tensor` of size [batch, height_in, width_in, channels].
factor: The subsampling factor.
scope: Optional variable_scope.
Returns:
output: A `Tensor` of size [batch, height_out, width_out, channels] with the
input, either intact (if factor == 1) or subsampled (if factor > 1).
"""
if factor == 1:
return inputs
else:
#return layers.max_pool2d(inputs, [1, 1], stride=factor, scope=scope)
return layers.pool(inputs, [2, 2, 2], stride=factor, scope=scope, data_format = 'NDHWC', pooling_type='MAX')
def teacher(input_images,
keep_prob,
is_training=True,
weight_decay=0.00004,
batch_norm_decay=0.997,
batch_norm_epsilon=0.001):
with tf.variable_scope("Teacher_model"):
net, endpoints = mobilenet_v2(inputs=input_images,
num_classes=FLAGS.num_class,
is_training=True,
spatial_squeeze=True,
scope='mobilenet_v2')
base_var_list = slim.get_model_variables()
for _ in range(2):
base_var_list.pop()
# feature & attention
t_g0 = endpoints["InvertedResidual_{}_{}".format(64, 3)]
t_at0 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g0), -1), axis=0, name='t_at0')
t_g1 = endpoints["InvertedResidual_{}_{}".format(96, 2)]
t_at1 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g1), -1), axis=0, name='t_at1')
part_feature = endpoints["InvertedResidual_{}_{}".format(160, 2)]
t_at2 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(part_feature), -1), axis=0, name='t_at3')
t_p_o = endpoints["InvertedResidual_{}_{}".format(320, 0)]
t_at3 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_p_o), -1), axis=0, name='t_at4')
object_feature = endpoints["Conv2d_8"]
t_at4 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(object_feature), -1), axis=0, name='t_at5')
t_g = (t_g0, t_g1, part_feature, object_feature)
t_at = (t_at0, t_at1, t_at2, t_at3, t_at4)
fc_obj = slim.max_pool2d(object_feature, (6, 8), scope="GMP1")
batch_norm_params = {
'center': True,
'scale': True,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
fc_obj = slim.conv2d(fc_obj,
M,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope='fc_obj')
fc_obj = tf.nn.dropout(fc_obj, keep_prob=keep_prob)
fc_obj = slim.flatten(fc_obj)
fc_part = slim.conv2d(part_feature,
M * k,
[1, 1],
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay)
)
fc_part = slim.max_pool2d(fc_part, (6, 8), scope="GMP2")
ft_list = tf.split(fc_part,
num_or_size_splits=FLAGS.num_class,
axis=-1)
cls_list = []
for i in range(M):
ft = tf.transpose(ft_list[i], [0, 1, 3, 2])
cls = layers_lib.pool(ft,
[1, 10],
"AVG")
cls = layers.flatten(cls)
cls_list.append(cls)
fc_ccp = tf.concat(cls_list, axis=-1)
fc_part = slim.conv2d(fc_part,
FLAGS.num_class,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope="fc_part")
fc_part = tf.nn.dropout(fc_part, keep_prob=keep_prob)
fc_part = slim.flatten(fc_part)
t_var_list = slim.get_model_variables()
return t_g, t_at, fc_obj, fc_part, fc_ccp, base_var_list, t_var_list
def student(input_images,
keep_prob,
is_training=True,
weight_decay=0.00004,
batch_norm_decay=0.997,
batch_norm_epsilon=0.001):
with tf.variable_scope("Student_model"):
net, endpoints = studentnet_v2.studentnet_v2(inputs=input_images,
num_classes=FLAGS.num_class,
is_training=True,
scope='student')
# feature & attention
s_g0 = endpoints["InvertedResidual_{}_{}".format(64, 1)]
s_at0 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(s_g0), -1), axis=0, name='s_at0')
s_g1 = endpoints["InvertedResidual_{}_{}".format(96, 1)]
s_at1 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(s_g1), -1), axis=0, name='s_at1')
s_g2 = endpoints["InvertedResidual_{}_{}".format(160, 1)]
s_at2 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(s_g2), -1), axis=0, name='s_at2')
s_g3 = endpoints["InvertedResidual_{}_{}".format(256, 0)]
s_at3 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(s_g3), -1), axis=0, name='s_at4')
s_g4 = endpoints["Conv2d_8"]
s_at4 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(s_g4), -1), axis=0, name='s_at5')
s_g = (s_g0, s_g1, s_g2, s_g4)
s_at = (s_at0, s_at1, s_at2, s_at3, s_at4)
s_part_feature = s_g2
s_object_feature = s_g4
base_var_list = slim.get_model_variables('Student_model/student')
batch_norm_params = {
'center': True,
'scale': True,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
# obj
s_fc_obj = slim.max_pool2d(s_object_feature, (6, 8), scope="s_GMP1")
s_fc_obj = slim.conv2d(s_fc_obj,
num_outputs=M,
kernel_size=[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope='s_fc_obj')
s_fc_obj = tf.nn.dropout(s_fc_obj, keep_prob=keep_prob)
s_fc_obj = layers.flatten(s_fc_obj)
s_fc_part = slim.conv2d(s_part_feature,
M * k,
[1, 1],
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay)
)
s_fc_part = slim.max_pool2d(s_fc_part, (6, 8), scope="s_GMP2")
s_ft_list = tf.split(s_fc_part,
num_or_size_splits=FLAGS.num_class,
axis=-1)
s_cls_list = []
for i in range(M):
s_ft = tf.transpose(s_ft_list[i], [0, 1, 3, 2])
s_cls = layers_lib.pool(s_ft,
[1, 10],
"AVG")
s_cls = layers.flatten(s_cls)
s_cls_list.append(s_cls)
s_fc_ccp = tf.concat(s_cls_list, axis=-1)
s_fc_part = slim.conv2d(s_fc_part,
FLAGS.num_class,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope="s_fc_part")
s_fc_part = tf.nn.dropout(s_fc_part, keep_prob=keep_prob)
s_fc_part = slim.flatten(s_fc_part)
s_fc_list = []
s_var_list = slim.get_model_variables('Student_model')
for var in s_var_list:
if var not in base_var_list:
s_fc_list.append(var)
# print('base_var_list',base_var_list)
# exit()
return s_g, s_at, s_fc_obj, s_fc_part, s_fc_ccp, base_var_list, s_var_list
def teacher(input_images,
keep_prob,
lambda_decay=FLAGS.lambda_decay,
is_training=True,
weight_decay=0.00004,
batch_norm_decay=0.99,
batch_norm_epsilon=0.001):
with tf.variable_scope("Teacher_model"):
net, endpoints = resnet.resnet_v2(inputs=input_images,
lambda_decay=lambda_decay,
num_classes=FLAGS.num_class,
is_training=True,
scope='resnet_v2_50')
# co_trained layers
var_scope = 'Teacher_model/resnet_v2_50/'
co_list_0 = slim.get_model_variables(var_scope + 'Conv2d_0')
# co_list_1 = slim.get_model_variables(var_scope +'InvertedResidual_16_')
# co_list_2 = slim.get_model_variables(var_scope +'InvertedResidual_24_')
t_co_list = co_list_0
base_var_list = slim.get_variables()
# for _ in range(2):
# base_var_list.pop()
lambda_c_list = slim.get_variables_by_name('lambda_c')
lambda_b_list = slim.get_variables_by_name('lambda_b')
t_lambda_list = lambda_c_list + lambda_b_list
# print(lambda_b_list)
# exit()
t_net_var_list =[]
for v in base_var_list:
if v not in t_co_list and v not in t_lambda_list:
t_net_var_list.append(v)
# feature & attention
t_g0 = endpoints["InvertedResidual_{}_{}".format(256, 2)]
t_at0 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g0), -1), axis=0, name='t_at0')
t_g1 = endpoints["InvertedResidual_{}_{}".format(512, 3)]
t_at1 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g1), -1), axis=0, name='t_at1')
part_feature = endpoints["InvertedResidual_{}_{}".format(1024, 5)]
t_at2 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(part_feature), -1), axis=0, name='t_at2')
object_feature = endpoints["InvertedResidual_{}_{}".format(2048, 2)]
t_at3 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(object_feature), -1), axis=0, name='t_at3')
# print(t_at1.get_shape().as_list())
# exit()
t_g = (t_g0, t_g1, part_feature, object_feature)
t_at = (t_at0, t_at1, t_at2, t_at3)
fc_obj = slim.max_pool2d(object_feature, (6, 8), scope="GMP1")
batch_norm_params = {
'center': True,
'scale': True,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
fc_obj = slim.conv2d(fc_obj,
M,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope='fc_obj')
fc_obj = tf.nn.dropout(fc_obj, keep_prob=keep_prob)
fc_obj = slim.flatten(fc_obj)
#
fc_part = slim.conv2d(part_feature,
M * k, #卷积核个数
[1, 1], #卷积核高宽
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm, # 标准化器设置为BN
normalizer_params=batch_norm_params,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay)
)
# print('part',fc_part.get_shape())
fc_part = slim.max_pool2d(fc_part, (12, 16), scope="GMP2")
ft_list = tf.split(fc_part,
num_or_size_splits=FLAGS.num_class,
axis=-1) #最后一维度（C）
cls_list = []
for i in range(M):
ft = tf.transpose(ft_list[i], [0, 1, 3, 2])
cls = layers_lib.pool(ft,
[1, 10],
"AVG")
cls = layers.flatten(cls)
cls_list.append(cls)
fc_ccp = tf.concat(cls_list, axis=-1) #cross_channel_pooling (N, M)
fc_part = slim.conv2d(fc_part,
FLAGS.num_class,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope="fc_part")
fc_part = tf.nn.dropout(fc_part, keep_prob=keep_prob)
fc_part = slim.flatten(fc_part)
t_var_list = slim.get_model_variables()
t_fc_var_list = []
for var in t_var_list:
if var not in base_var_list:
t_fc_var_list.append(var)
return t_g, t_at, fc_obj, fc_part, fc_ccp, t_co_list, t_net_var_list, t_fc_var_list, t_lambda_list, t_var_list
def dfb(input_images,
keep_prob,
is_training=True,
weight_decay=5e-5,
batch_norm_decay=0.99,
batch_norm_epsilon=0.001):
with tf.variable_scope("Teacher_model"):
net, endpoints = resnet_v2(inputs=input_images,
num_classes=M,
is_training=True,
scope='resnet_v2')
base_var_list = slim.get_model_variables('Teacher_model/resnet_v2')
part_feature = endpoints["InvertedResidual_{}_{}".format(1024, 3)]
object_feature = endpoints["InvertedResidual_{}_{}".format(1024, 5)]
object_feature_h = object_feature.get_shape().as_list()[1]
object_feature_w = object_feature.get_shape().as_list()[2]
fc_obj = slim.max_pool2d(object_feature, (object_feature_h, object_feature_w), scope="GMP1")
batch_norm_params = {
'center': True,
'scale': True,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
fc_obj = slim.conv2d(fc_obj,
M,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope='fc_obj')
fc_obj = tf.nn.dropout(fc_obj, keep_prob=keep_prob)
fc_obj = slim.flatten(fc_obj)
fc_part = slim.conv2d(part_feature,
M * k, #卷积核个数
[1, 1], #卷积核高宽
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm, # 标准化器设置为BN
normalizer_params=batch_norm_params,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay)
)
fc_part_h = fc_part.get_shape().as_list()[1]
fc_part_w = fc_part.get_shape().as_list()[2]
fc_part = slim.max_pool2d(fc_part, (fc_part_h, fc_part_w), scope="GMP2")
ft_list = tf.split(fc_part,
num_or_size_splits=M,
axis=-1) #最后一维度（C）
cls_list = []
for i in range(M):
ft = tf.transpose(ft_list[i], [0, 1, 3, 2])
cls = layers_lib.pool(ft,
[1, k],
"AVG")
cls = layers.flatten(cls)
cls_list.append(cls)
fc_ccp = tf.concat(cls_list, axis=-1) #cross_channel_pooling (N, M)
fc_part = slim.conv2d(fc_part,
M,
[1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope="fc_part")
fc_part = tf.nn.dropout(fc_part, keep_prob=keep_prob)
fc_part = slim.flatten(fc_part)
t_var_list = slim.get_model_variables()
return fc_obj, fc_part, fc_ccp, base_var_list, t_var_list
def teacher(input_images,
keep_prob,
is_training=True,
weight_decay=5e-5,
batch_norm_decay=0.99,
batch_norm_epsilon=0.001):
with tf.variable_scope("Teacher_model"):
net, endpoints = resnet_v2(inputs=input_images,
num_classes=M,
is_training=True,
scope='resnet_v2')
# co_trained layers
var_scope = 'Teacher_model/resnet_v2/'
co_list_0 = slim.get_model_variables(var_scope + 'Conv2d_0')
# co_list_1 = slim.get_model_variables(var_scope +'InvertedResidual_16_0/conv')
# co_list_2 = slim.get_model_variables(var_scope +'InvertedResidual_24_')
t_co_list = co_list_0
base_var_list = slim.get_model_variables('Teacher_model/resnet_v2')
# feature & attention
t_g0 = endpoints["InvertedResidual_{}_{}".format(256, 2)]
t_at0 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g0), -1),
axis=0,
name='t_at0')
t_g1 = endpoints["InvertedResidual_{}_{}".format(512, 3)]
t_at1 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g1), -1),
axis=0,
name='t_at1')
part_feature = endpoints["InvertedResidual_{}_{}".format(1024, 3)]
t_at2 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(part_feature), -1),
axis=0,
name='t_at2')
t_g3 = endpoints["InvertedResidual_{}_{}".format(1024, 4)]
t_at3 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(t_g3), -1),
axis=0,
name='t_at3')
object_feature = endpoints["InvertedResidual_{}_{}".format(1024, 5)]
t_at4 = tf.nn.l2_normalize(tf.reduce_sum(tf.square(object_feature),
-1),
axis=0,
name='t_at4')
t_g = (t_g0, t_g1, part_feature, object_feature)
t_at = (t_at0, t_at1, t_at2, t_at3, t_at4)
object_feature_h = object_feature.get_shape().as_list()[1]
object_feature_w = object_feature.get_shape().as_list()[2]
fc_obj = slim.max_pool2d(object_feature,
(object_feature_h, object_feature_w),
scope="GMP1")
batch_norm_params = {
'center': True,
'scale': True,
'decay': batch_norm_decay,
'epsilon': batch_norm_epsilon,
}
fc_obj = slim.conv2d(
fc_obj,
M, [1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope='fc_obj')
fc_obj = tf.nn.dropout(fc_obj, keep_prob=keep_prob)
fc_obj = slim.flatten(fc_obj)
fc_part = slim.conv2d(
part_feature,
M * k, #卷积核个数
[1, 1], #卷积核高宽
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm, # 标准化器设置为BN
normalizer_params=batch_norm_params,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay))
fc_part_h = fc_part.get_shape().as_list()[1]
fc_part_w = fc_part.get_shape().as_list()[2]
fc_part = slim.max_pool2d(fc_part, (fc_part_h, fc_part_w),
scope="GMP2")
ft_list = tf.split(fc_part, num_or_size_splits=M, axis=-1) #最后一维度（C）
cls_list = []
for i in range(M):
ft = tf.transpose(ft_list[i], [0, 1, 3, 2])
cls = layers_lib.pool(ft, [1, k], "AVG")
cls = layers.flatten(cls)
cls_list.append(cls)
fc_ccp = tf.concat(cls_list, axis=-1) #cross_channel_pooling (N, M)
fc_part = slim.conv2d(
fc_part,
M, [1, 1],
activation_fn=None,
weights_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
biases_regularizer=tf.contrib.layers.l2_regularizer(weight_decay),
scope="fc_part")
fc_part = tf.nn.dropout(fc_part, keep_prob=keep_prob)
fc_part = slim.flatten(fc_part)
t_var_list = slim.get_model_variables()
return t_co_list, t_g, t_at, fc_obj, fc_part, fc_ccp, base_var_list, t_var_list